This study investigated the influence of caption descriptions used during training on AI-generated Kawaii fashion styles. By controlling caption content while fixing visual data and training conditions, it was shown that explicit
linguistic references to Harajuku Kawaii improve stylistic coherence and balance in generated images. Models trained with explicit Kawaii-related captions successfully integrated decorative Kawaii elements with CHANEL-like elegance, whereas models without such captions often produced unstable or excessive expressions. These results demonstrate that caption design is a critical factor in shaping stylistic interpretation in image-generation AI.
Peeraya SRIPIAN, Keiko MIYATAKE, Tippporn LAOHAKANGVALVIT, and Michiko OHKURA
